Machine learning and operation research based method for promotion optimization of products with no price elasticity history

Highlights：

• Many leading e-commerce retailers adopt a consistent pricing strategy to build customer trust and promote just a small portion of their catalog every week.

• Promotion optimization for consistent pricing retailers is a challenging problem, as they need to decide which products to promote with no historical price elasticity information for the candidate products.

• This research introduces a novel approach for predicting product price elasticity impact with no historical price elasticity information, to enable promotion optimization by retailers who adopt a consistent pricing strategy.

• Our approach combines the commonly used operation research based log–log demand model and the GBM regression algorithm incorporating semantic product similarity related features to predict price elasticity impact for products with no historical price elasticity information.

• We employ a pessimistic prediction interval measure to accelerate the learning, and incorporate the prediction uncertainty when ranking candidate products for reducing the probability of recommending low impact promotions due to model prediction uncertainty.

• We introduce a set of features based on consumers browsing clickstream in an e-commerce website, in order to capture the fine granularity of product similarity and incorporate it into the prediction model to further improve price elasticity impact learning.

• An evaluation of the proposed approach using 6 months of real world extensive dataset collected from an European online department store, during two independent periods demonstrates the superiority of the proposed approach versus the commonly used operation research based log–log demand model for this use case for both price elasticity impact prediction and promotion revenue optimization. Furthermore, we show that employing the pessimistic prediction interval during the training period reduces the number of promotions required to predict the price elasticity of product with no price elasticity history.

摘要

•Many leading e-commerce retailers adopt a consistent pricing strategy to build customer trust and promote just a small portion of their catalog every week.•Promotion optimization for consistent pricing retailers is a challenging problem, as they need to decide which products to promote with no historical price elasticity information for the candidate products.•This research introduces a novel approach for predicting product price elasticity impact with no historical price elasticity information, to enable promotion optimization by retailers who adopt a consistent pricing strategy.•Our approach combines the commonly used operation research based log–log demand model and the GBM regression algorithm incorporating semantic product similarity related features to predict price elasticity impact for products with no historical price elasticity information.•We employ a pessimistic prediction interval measure to accelerate the learning, and incorporate the prediction uncertainty when ranking candidate products for reducing the probability of recommending low impact promotions due to model prediction uncertainty.•We introduce a set of features based on consumers browsing clickstream in an e-commerce website, in order to capture the fine granularity of product similarity and incorporate it into the prediction model to further improve price elasticity impact learning.•An evaluation of the proposed approach using 6 months of real world extensive dataset collected from an European online department store, during two independent periods demonstrates the superiority of the proposed approach versus the commonly used operation research based log–log demand model for this use case for both price elasticity impact prediction and promotion revenue optimization. Furthermore, we show that employing the pessimistic prediction interval during the training period reduces the number of promotions required to predict the price elasticity of product with no price elasticity history.